GPT Image 2.6: how to choose between the three image models

The three models are exposed through the production image catalog. gpt-image-2.6 is the general starting point for a straightforward subject or scene. gpt-image-2.6-sunburst is the preferred starting point for a person, several references, a detailed scene or a result where visual continuity matters. gpt-image-2.6-flare is the preferred starting point for posters, packaging, menus, diagrams and other compositions where readable text or a dense layout is central. These are Sparkles routing roles based on current observations, not a claim that a model is guaranteed to win every prompt.

Start with the goal rather than the model name.

The automatic mode uses these signals and falls back to an available catalog model. If the first result is not stable enough, switch to a fixed model and repeat the same request so you can compare like with like.

Sparkles exposes the image controls that are currently validated by the production catalog: low, medium, high and auto quality; aspect-ratio choices; auto, opaque or transparent backgrounds; and PNG, JPEG or WebP output. You can attach reference images, ask for a particular number of images and keep a series of related results in one workflow. Mention the words that must appear in the picture, then inspect them at full size. These controls describe the current product contract and may change when the catalog changes.

Choose automatic when you are exploring, when the task changes from turn to turn, or when you want the assistant to route a poster, reference-heavy scene or ordinary subject for you. Choose a fixed model when you are testing a prompt, comparing revisions, building a series or need to reproduce a known direction. The selected preference applies to new work. A submitted task keeps the model and parameters that were frozen when it started, so changing the setting does not rewrite a finished or queued image.

A reference image supplies visual evidence; it does not make every detail exact. Say which reference controls identity, which one only supplies atmosphere and what may change. For packaging, logos, ingredients, dates and small type, compare the output with approved artwork or a source document. Transparent output is useful for compositing, but preview it against both light and dark backgrounds to find unwanted edges. Keep important facts as editable text in the final layout when the image will be published.

Image generation uses Credits and the final charge depends on the selected model, quality, dimensions and references. The pricing page is the current source of truth; there is no single flat per-image promise. The 2.6 catalog currently exposes low, medium, high and auto. Do not select older xhigh or max values for these models unless the catalog explicitly adds them. A successful provider request is evidence that the route works, not a guarantee of a particular visual result.

Try a small request before a long series:

Create one 4:5 image of [subject] for [purpose]. Preserve [identity details] from my reference. Use [model preference: automatic / gpt-image-2.6 / gpt-image-2.6-sunburst / gpt-image-2.6-flare], medium quality and a transparent background only if the edges will be composited. Keep [exact words] short and legible. Tell me what needs checking before I use the result.

Review the subject, text, edges and composition at full size. Then keep the model fixed while you revise one issue at a time.